LEAR: LLM-Driven Evolution of Agent-Based Rules

This study investigates the feasibility and effectiveness of integrating Large Language Models (LLMs) as mutation operators within Genetic Programming (GP) frameworks so as to evolve agent behaviors in multi-agent systems (MAS) and provide benchmarks that evaluate the efficacy of LLM-generated code in multi-agent domains. Our approach leverages the sophisticated code-generation capabilities of LLMs to introduce semantically meaningful variations during the evolutionary process. Specifically, we explore and systematically compare these different prompting strategies: zero-shot, one-shot, and two-shot prompting as well as prompting the generation of commented code to assess their impact on the quality of evolved agent behaviors. Additionally, we propose a novel methodology where evolution operates at a higher abstraction level by mutating pseudocode representations of agent behaviors, subsequently converting them into executable code through another LLM-mediated step. This strategy capitalizes on the extensive natural language training data of LLMs, potentially enabling the discovery of more innovative solutions. Our results indicate that LLM-driven mutation with comment generation enhances agent performance while mutating pseudocode representations yields reduced performance. This research contributes valuable insights regarding the integration of LLM-driven GP techniques into MAS, highlighting both the potential and limitations of these approaches. All code is open-sourced at https://github.com/can-gurkan/LEAR.